Scholarlink

<p>Continuous casting slabs are vital intermediates in steel production, necessitating real-time and accurate detection. At present, the programmable logic controller (PLC)-based continuous casting slab detection system used in the production site has limitations such as difficulty in realizing continuous position detection, sensor false triggering, aging, and even failure. Meanwhile, conventional vision-based target detection algorithms struggle with the complex physical environment of continuous casting sites, characterized by the complex imaging conditions, the drastic scale variations of targets due to camera layouts, and the frequent occlusions by heavy machinery. Moreover, processing high-resolution video streams from multiple cameras generates massive data throughput, creating immense computational pressure. Given these limitations, this paper proposes a system-level customized optimization algorithm named ICCS-YOLOv8 (Images of Continuous Casting Slabs—You Only Look Once version 8). Firstly, a partial transformer block (PTBlock) was designed to enhance adaptability to slab geometric variations and mitigate representational attenuation of small target features. Secondly, the selective boundary aggregation (SBA) was integrated into the feature pyramid network (FPN) structure to achieve bidirectional integration of features across different resolutions. Thirdly, a dedicated high-resolution detection layer was incorporated, and the parameter scale of the detection head was optimized through shared convolutions. Experimental results on a self-constructed proprietary dataset demonstrate a 97.4% mean average precision (mAP)@50 with a computational complexity of 13.1 giga floating-point operations (GFLOPs). Compared with the YOLOv8n baseline, the proposed algorithm improves detection accuracy by 2.3% while maintaining a favorable balance between detection accuracy and model complexity. Especially, the deployment experiment shows that the ICCS-YOLOv8 satisfies the production requirements of on-site detection speed ≥ 15 frames per second (FPS), thereby validating its potential for rapid and accurate continuous casting slab detection within complex industrial backgrounds. The code is available at <a href="https://github.com/llgu1stumail/ICCS-YOLOv8">https://github.com/llgu1stumail/ICCS-YOLOv8</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhanced YOLOv8 with global attention for precise continuous casting slab detection

  • 期刊论文

  • 发表日期：

    2026年3月5日

查看全文
  • Le Liu,
  • Linglong Gu,
  • Ziyang Li,
  • Yiming Fang

摘要

Continuous casting slabs are vital intermediates in steel production, necessitating real-time and accurate detection. At present, the programmable logic controller (PLC)-based continuous casting slab detection system used in the production site has limitations such as difficulty in realizing continuous position detection, sensor false triggering, aging, and even failure. Meanwhile, conventional vision-based target detection algorithms struggle with the complex physical environment of continuous casting sites, characterized by the complex imaging conditions, the drastic scale variations of targets due to camera layouts, and the frequent occlusions by heavy machinery. Moreover, processing high-resolution video streams from multiple cameras generates massive data throughput, creating immense computational pressure. Given these limitations, this paper proposes a system-level customized optimization algorithm named ICCS-YOLOv8 (Images of Continuous Casting Slabs—You Only Look Once version 8). Firstly, a partial transformer block (PTBlock) was designed to enhance adaptability to slab geometric variations and mitigate representational attenuation of small target features. Secondly, the selective boundary aggregation (SBA) was integrated into the feature pyramid network (FPN) structure to achieve bidirectional integration of features across different resolutions. Thirdly, a dedicated high-resolution detection layer was incorporated, and the parameter scale of the detection head was optimized through shared convolutions. Experimental results on a self-constructed proprietary dataset demonstrate a 97.4% mean average precision (mAP)@50 with a computational complexity of 13.1 giga floating-point operations (GFLOPs). Compared with the YOLOv8n baseline, the proposed algorithm improves detection accuracy by 2.3% while maintaining a favorable balance between detection accuracy and model complexity. Especially, the deployment experiment shows that the ICCS-YOLOv8 satisfies the production requirements of on-site detection speed ≥ 15 frames per second (FPS), thereby validating its potential for rapid and accurate continuous casting slab detection within complex industrial backgrounds. The code is available at https://github.com/llgu1stumail/ICCS-YOLOv8.